---
format: "aidr-story-markdown/v1"
id: "3e610ae2436c7a510ff0d1e113010c2277b2940adafe1e723df452e44aeb117e"
canonical_url: "https://aidr.today/3e610ae2?lang=en"
title: "PromptArmor Benchmark Breaches Every Frontier AI Model via Hidden Instructions"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-18T19:28:58.000Z"
category: "Research"
topics: ["safety"]
source_urls: ["https://marketbrief.now/ai/promptarmor-benchmark-breaches-every-frontier-ai-model-via-hidden-instru-68a57df7","https://huggingnews.com/ai/promptarmor-benchmark-breaches-every-frontier-ai-model-via-hidden-instru-68a57df7"]
summary: "Hidden-instruction attacks were successful in 34.9% of tests across 21 AI models, according to a new benchmark called FrontierBreach released by PromptArmor. The benchmark, which focuses on agentic workflows in enterprise environments, resulted in the breach of every frontier model tested. The attacks involve hiding invisible text within emails or documents that mislead AI agents into running attacker code or exfiltrating data. Among the models tested, Grok 4.5 fell for 91.4% of attacks, whereas Gemini Argon was breached in 4.3% of cases. PromptArmor also found that attacks hidden inside agent skills worked in 45.5% of tests."
---

# PromptArmor Benchmark Breaches Every Frontier AI Model via Hidden Instructions

> [Open the canonical story](<https://aidr.today/3e610ae2?lang=en>)

**Published:** 2026-11-18T19:28:58.000Z
**Category:** Research
**Topics:** safety

## Summary

Hidden\-instruction attacks were successful in 34\.9% of tests across 21 AI models, according to a new benchmark called FrontierBreach released by PromptArmor\. The benchmark, which focuses on agentic workflows in enterprise environments, resulted in the breach of every frontier model tested\. The attacks involve hiding invisible text within emails or documents that mislead AI agents into running attacker code or exfiltrating data\. Among the models tested, Grok 4\.5 fell for 91\.4% of attacks, whereas Gemini Argon was breached in 4\.3% of cases\. PromptArmor also found that attacks hidden inside agent skills worked in 45\.5% of tests\.

## Sources

- [Story source](<https://marketbrief.now/ai/promptarmor-benchmark-breaches-every-frontier-ai-model-via-hidden-instru-68a57df7>)
- [Story source](<https://huggingnews.com/ai/promptarmor-benchmark-breaches-every-frontier-ai-model-via-hidden-instru-68a57df7>)

